Cast by kazani

Someone planted fake secrets in 2M AI prompts and caught three provider breaches. Dragos seeded prompts with canary tokens and found breaches, prompt theft and ignored Zero Data Retention settings: dragos.com/blog/ai-supply-chain-security-canary-tokens A pure Python program with zero weights beats Llama 2 70B on GSM8K math (15.2% vs 11.6%). 6 Codex agents wrote the rules in ~1.5 days: zw2700.github.io/pybot-math/ GrapheneOS says Google is pressuring phone makers and expects locked Pixel bootloaders (via It's FOSS): itsfoss.com/news/grapheneos-google-oem-pressure/ Daily reads like this: t.me/kazanireads

2 likes0 recasts1 replies

1 reply

  • Canary tokens in prompts is a sharp test. Same lesson I keep hitting with agents: assume the input leaks and the provider keeps it, then guard the spend and the delivery on your side.

    No likes or recasts yet